Winds and Ocean Currents. Latent Heat Transport 580 cal/g Surface wind Surface wind.
Examination of Errors in Near-Surface Temperature and Wind...
Transcript of Examination of Errors in Near-Surface Temperature and Wind...
Examination of Errors in Near-Surface Temperature and Wind from WRFNumerical Simulations in Regions of Complex Terrain
HAILING ZHANG AND ZHAOXIA PU
Department of Atmospheric Sciences, University of Utah, Salt Lake City, Utah
XUEBO ZHANG
Computational Engineering and Science, University of Utah, Salt Lake City, Utah
(Manuscript received 11 October 2012, in final form 4 January 2013)
ABSTRACT
The performance of an advanced research version of theWeather Research and ForecastingModel (WRF)
in predicting near-surface atmospheric temperature and wind conditions under various terrain and weather
regimes is examined. Verification of 2-m temperature and 10-m wind speed and direction against surface
Mesonet observations is conducted. Three individual events under strong synoptic forcings (i.e., a frontal
system, a low-level jet, and a persistent inversion) are first evaluated. It is found that theWRFmodel is able to
reproduce these weather phenomena reasonably well. Forecasts of near-surface variables in flat terrain
generally agree well with observations, but errors also occur, depending on the predictability of the lower-
atmospheric boundary layer. In complex terrain, forecasts not only suffer from the model’s inability to re-
produce accurate atmospheric conditions in the lower atmosphere but also struggle with representative issues
due tomismatches between themodel and the actual terrain. In addition, surface forecasts at finer resolutions
do not always outperform those at coarser resolutions. Increasing the vertical resolution may not help predict
the near-surface variables, although it does improve the forecasts of the structure of mesoscale weather
phenomena. A statistical analysis is also performed for 120 forecasts during a 1-month period to further
investigate forecast error characteristics in complex terrain. Results illustrate that forecast errors in near-
surface variables depend strongly on the diurnal variation in surface conditions, especially when synoptic
forcing is weak. Under strong synoptic forcing, the diurnal patterns in the errors break down, while the flow-
dependent errors are clearly shown.
1. Introduction
The near-surface atmosphere, namely, the bottom
10%of the atmospheric boundary layer, is unique due to
its direct interaction with the earth’s surface (Stull 1988).
For instance, near-surface temperature is characterized
by diurnal variation, with a maximum at local afternoon
and a minimum at local midnight. This is very different
from the free atmosphere in which temperature shows
little diurnal variation. Turbulence causes the wind field
in the atmospheric boundary layer (ABL) and the near-
surface to behave differently from that in the free at-
mosphere because the ABL transports momentum,
heat, and moisture between the earth’s surface and the
air above. Due to its unique features, accurate fore-
casts of near-surface atmospheric conditions are very
important in many applications such as wind energy,
agriculture, aviation, and fire weather forecasts. How-
ever, difficulties in forecasting near-surface variables
such as temperature and wind have long been recog-
nized and studied (Hanna and Yang 2001; Zhang and
Zheng 2004).
To accurately simulate near-surface atmospheric
conditions, several factors must be represented properly
in numerical models. These include land use, topogra-
phy, surface heat flux transport, and various character-
istics of the lower atmosphere (Lee et al. 1989; Wolyn
and McKee 1989; Shafran et al. 2000; Cheng and
Steenburgh 2005). Thus, the accurate simulation of
near-surface atmospheric diurnal variation is one of the
most important and difficult tasks in numerical weather
Corresponding author address: Dr. Zhaoxia Pu, Dept. of At-
mospheric Sciences, Rm. 819, University of Utah, 135 S 1460 E,
Salt Lake City, UT 84112.
E-mail: [email protected]
JUNE 2013 ZHANG ET AL . 893
DOI: 10.1175/WAF-D-12-00109.1
� 2013 American Meteorological Society
prediction (NWP). Owing to our limited understanding
of near-surface atmospheric processes and the uncer-
tainties in model physics parameterizations, a compre-
hensive verification of the NWP models’ performance in
forecasting near-surface variables becomes a necessary
step for model improvement.
Hanna and Yang (2001) found that the uncertainties
regarding wind speed and direction in the lower at-
mosphere are primarily due to random turbulent pro-
cesses that were not appropriately represented in the
models, as well as errors in subgrid terrain and land
use. They also argued that the models tend to un-
derestimate the vertical temperature gradients in the
lowest 100 m during the nighttime. Thus, the simulated
boundary layer stability is not as strong as the ob-
served. Considering the different capabilities of plan-
etary boundary layer (PBL) parameterization schemes
to reproduce atmospheric structures in the lowest few
kilometers, Zhang and Zheng (2004) tested the per-
formances of different PBL schemes in simulating
near-surface temperature and wind speed and di-
rection. Their results revealed that the model could
reproduce diurnal variations in surface temperature
and wind direction. However, all the boundary layer
schemes underestimated (overestimated) wind speeds
during the daytime (nighttime). Their study was con-
ducted over the central United States during the summer,
where little organized convection and topographical
forcing was present.
The problem becomes more complicated in complex
terrain. Liu et al. (2008) conducted an interrange
comparison of the model analyses and forecasts of five
U.S. Army test and evaluation command ranges over
a 5-yr period. They concluded that forecast errors vary
from range to range and season to season. They also
found that larger errors are typically associated with
complex terrain. Zhong and Fast (2003) compared
three mesoscale numerical models and evaluated the
simulations over the Salt Lake Valley for cases influ-
enced by both weak and strong synoptic scenarios.
They found a cold bias in the valley extending from the
surface to the top of the atmosphere. The simulated
nocturnal inversion was much weaker than the ob-
served. There were significant errors in wind forecasts
even under strong synoptic forcing. Hart et al. (2005)
validated surface forecasts over mountainous terrain
during wintertime. They employed the fifth-generation
Pennsylvania State University–National Center for
Atmospheric Research Mesoscale Model (MM5) at
high resolution with three nested domains at 36-, 12-,
and 4-km horizontal grid spacing. Simulation results
did show improved wind and precipitation forecasts in
the 4-km horizontal grid-spacing domain, compared
with those at the 12- and 36-km domains. However,
temperature forecasts did not benefit from the high-
resolution simulation. It is noteworthy that although
the model properly simulated the persistent nocturnal
cold-air pool along with the better-resolved orography
at higher resolution, it still did not improve the 2-m
temperature forecasts. Apparently, forecasting sur-
face conditions in complex terrain is a challenging
problem.
Mass et al. (2002) presented an objective multiyear
verification of the University of Washington real-time
MM5 forecasts. In their study, triple one-way nested
domains at 36-, 12-, and 4-km horizontal grid spacings
were used and the forecasts of near-surface atmospheric
conditions (i.e., 2-m temperature, 10-m wind direction
and speed, precipitation, and sea level pressure) from all
three domains were verified at observation locations in
western Washington State. Their results suggested that
the forecasts benefited significantly from decreasing the
grid spacing from 36 to 12 km. However, little im-
provement was found with further reduction in grid
spacing from 12 to 4 km. These results are in contrast
with the conclusions of some other studies. For instance,
Rife andDavis (2005) suggested that the gains in forecast
accuracy from finer grid spacing are generally incre-
mental. With a regional climate simulation over complex
terrain, Leung and Qian (2003) found that a higher-
resolution simulation improves not only the spatial dis-
tribution and regional mean precipitation during summer
but also snowpack during winter. However, they also
commented that the accuracy of snow simulation is lim-
ited by factors such as deficiencies in the land surface
model or biases in other model variables. The disagree-
ment between these different studies further indicates the
complexity of numerical prediction over complex terrain.
Nevertheless, most of these previous studies emphasized
the verification of synoptic cases at large and mesoscales.
Little attention has been paid to the assessment of near-
surface atmospheric conditions.
In this study, we attempt to assess the accuracy of the
near-surface atmospheric conditions, specifically the 2-m
temperature and 10-m wind speed and direction, in nu-
merical simulations produced by the Weather Research
and Forecasting Model (WRF). In particular, version
3.3 of an Advanced Research version of the WRF
(ARW; Skamarock et al. 2008) is used for three typical
severe weather events (i.e., a low-level jet, a cold front,
and a wintertime persistent inversion) over the south-
ern Great Plains (SGP) and the Intermountain West of
the United States. Our purposes are not only to ex-
amine the ability of the ARW to predict near-surface
atmospheric conditions, but also to compare the pre-
dictability of near-surface conditions in flat and
894 WEATHER AND FORECAST ING VOLUME 28
complex terrain. The sensitivity of numerical forecasts
of near-surface atmospheric conditions to various PBL
schemes and model resolutions is also investigated. In
addition, forecasts during a 1-month period are evalu-
ated to further reveal the characteristics of the forecast
errors of near-surface variables under different synoptic
forcings in complex terrain.
This paper is organized as follows. Section 2 briefly
describes three synoptic events in two individual cases,
as well as the numerical simulations and verification
methods. Sections 3 and 4 detail the simulation and
verification results for the three synoptic events. Error
characteristics of the near-surface variables are also
evaluated. Section 5 examines the sensitivity of numer-
ical simulations of near-surface atmospheric conditions
to various PBL parameterization schemes and model
vertical resolutions. Section 6 characterizes the errors in
near-surface variables statistically with forecasts during
a 1-month period. Section 7 summarizes the results and
offers several concluding remarks.
2. Description of cases, numerical simulations, andverification methods
a. Cases
1) 1–3 JUNE 2008: A FRONTAL SYSTEM AND
A LOW-LEVEL JET
There are two events of interest during 1–3 June 2008:
a front evolved over the north-central United States and
a nocturnal low-level jet occurred over the SGP. Surface
maps (not shown) show that a cold front initially located
north of North Dakota at 1200 UTC 1 June entered
North Dakota at 1500 UTC 1 June. It arrived in South
Dakota at 0100 UTC 2 June, then changed to a station-
ary front at 1200 UTC 2 June and evolved into a cold
front again as it moved southward. A temperature gra-
dient of 78C along with wind direction changes were
found between the two stations closest to the front on
both sides, with a southwest wind on the south side and
a northeast wind on the north side.
A low-level jet was dominant from the surface up to
1800 m above ground level (AGL) over the entire SGP.
It influenced near-surface conditions by interacting with
the surface and the lower atmosphere. Radar wind
profiles in Jayton, Texas, from the National Oceanic and
Atmospheric Administration (NOAA) Profiler Net-
work indicated two periods of evidently greater wind
speed during 1–3 June (see Fig. 4a in section 3a). One
was between 0200 and 1400 UTC 1 June and the other
was between 0200 and 1400 UTC 2 June.
2) 1–3DECEMBER 2010: A PERSISTENT INVERSION
A persistent inversion began on 29 November 2010
and was maintained over the Salt Lake Valley, Utah, for
seven successive days. During this period, extremely
strong surface cooling occurred during the night of 1–2
December, accompanied by a clear sky and a strong
temperature inversion layer aloft. Very low tempera-
tures were observed in the Salt Lake Valley and the
adjacent mountains. A high pressure system controlled
the area and helped build and maintain the persistent
cold air pools during this period through downward air
motion. The strong inversion layer extended from the
surface up to 2000 m during this time.
b. A brief description of numerical simulations
Numerical experiments are conducted to simulate the
aforementioned cases using the ARW with one-way
nested domains. The initial and boundary conditions
are derived from theNational Centers for Environmental
Prediction’s (NCEP’s) Northern American Mesoscale
(NAM) model analysis by WRF preprocessing. A to-
pography dataset at 30 arc-second (about 1000 m) reso-
lution and an update land-use dataset with 27 land-use
categories (instead of the 24 land-use categories provided
by theWRFversion 3.3 release) from theU.S.Geological
Survey (USGS) are used in order to ensuremore accurate
surface conditions, especially for playa and desert regions
in the western United States. The Noah land surface
model is used because it predicts the land states, such as
TABLE 1. Configurations of numerical simulations.
Case 1–3 Jun 2008 1–3 Dec 2010 Fall 2011 (15 Sep–14 Oct)
No. of domains 3 3 4
Horizontal grid spacing (km) 27, 9, 3 12, 4, 1.33 30, 10, 3.33, 1.11
Microphysics scheme WSM 6 WSM 6 Purdue Lin
PBL scheme YSU MYJ YSU
Cumulus scheme Kain–Fritsch (not applied for grid spacing 9 km)
Land surface scheme Noah
Longwave radiation Rapid Radiative Transfer Model
Shortwave radiation Dudhia
JUNE 2013 ZHANG ET AL . 895
surface temperature and soil moisture and temperature,
in each layer with time. Table 1 lists the configurations
of horizontal resolutions, model domains, and physical
schemes used for each simulation in this study. Since
PBL parameterization schemes contain key physical
factors that strongly influence the predictability of
near-surface atmospheric conditions, the PBL scheme
used for each individual simulation (i.e., the control
simulation) is chosen from the sensitivity studies (as
described in section 5).
c. Verification methods
1) SYNOPTIC VERIFICATION
Verification is first conducted to evaluate the accuracy
of the numerical simulation of each synoptic event.
Simulation results are compared with available obser-
vations and analyses.
2) VERIFICATION OF NEAR-SURFACE
ATMOSPHERIC CONDITIONS
The major emphasis of this study is on characterizing
errors in the near-surface atmosphere. To quantify these
errors, we use surface Mesonet observations (Horel
et al. 2002) to verify themodel’s performance in terms of
the near-surface variables, namely, 2-m temperature
and 10-m wind speed and direction. According to Horel
et al. (2002), quality control algorithms and data moni-
toring programs are performed for all available data.
The quality-controlled data are then made available
hourly with quality flags. In this study, only those ob-
servations with a quality flag of ‘‘OK’’ (the highest
quality) are used for verification.
Since there is case-by-case variation in near-surface
atmospheric conditions due to various synoptic systems
and terrain, verification of near-surface atmospheric
FIG. 1. Locations of model domains for numerical simulations: (a) 0000 UTC 1 Jun–
0000 UTC 3 Jun 2008 and (b) 0000 UTC 1Dec–0000 UTC 3Dec 2010. Shaded contours denote
the terrain heights.
896 WEATHER AND FORECAST ING VOLUME 28
conditions is performed for each synoptic event. Rep-
resentation errors due to discrepancies between the
model and actual terrain heights are commonly present
in the forecasts of surface variables. Therefore, model
terrain heights are compared against actual terrain
heights for each case to examine the representative
errors. Model performance is then checked for each
case for each variable over time. In this study, we use
variable mean, mean absolute error (MAE), and bias
error (BE) of 2-m temperature and 10-m wind speed
and direction against observations to characterize the
errors in numerical simulations. We also calculate
time-averaged mean absolute errors (TMAEs or cu-
mulative MAEs) to average the MAEs over the whole
simulation period. Because observational errors in
wind direction are usually larger at lower wind speeds,
only those observations with wind speeds greater than
1.5 m s21 are used to verify wind direction. To verify
the model simulation, simulation results at model grid
points are interpolated to observation locations using
a bilinear method. The statistical calculations are as
follow:
MAE51
n�n
i51
jFi 2Oij
BE51
n�n
i51
(Fi 2Oi)
TMAE51
n
1
m�m
t51�n
i51
jFit 2Oitj ,
where i denotes the ith observation, t denotes the ob-
servation time, Oi represents the value of the observa-
tion at the ith location, Fi denotes the forecast value
interpolated to that observation location, n is the total
number of stations, and m represents the total number
of times used to calculate TMAE.
3. Simulation and verification: 1–3 June 2008
Triple-level, one-way nested domains (D01, D02, and
D03 in Fig. 1a) with 27-, 9-, and 3-km horizontal grid
FIG. 2. Weather maps at 850 hPa valid at 0000 UTC 2 Jun 2008.
(a) Geopotential heights (contour interval is 30 m) and wind barbs,
and (b) temperature (contour interval is 28C) with wind barbs.
Black contour lines and wind barbs represent observations from
the upper-level observation network. Blue contour lines and wind
barbs denote model simulations (48-h forecast). The thick black
curve in (b) marks the cold front.
FIG. 3. Horizontal wind speeds (contour interval is 4 m s21) valid at 0900 UTC 2 Jun 2008 at 850 hPa: (a) NARR
reanalysis and (b) 33-h model forecast from the 3-km domain. Wind speeds greater than 12 m s21 are shaded.
JUNE 2013 ZHANG ET AL . 897
spacings (hereafter referred to as the 27-, 9-, and 3-km
domains, respectively) are utilized in this simulation.
There are 37 vertical levels from the surface up to
50 hPa. The innermost domain (i.e., the 3-km domain)
focuses on the two weather systems of interest over
the north-central United States and the SGP. Themodel
is initialized at 0000 UTC 31 May, 24 h ahead of the
verification.
Physical parameterization options, as listed in Table 1,
include the WRF single-moment six-class microphysics
scheme (WSM6; Hong and Lim 2006), the Yonsei Uni-
versity (YSU) PBL scheme (Hong and Pan 1996), the
Kain–Fritsch cumulus parameterization scheme (Kain
and Fritsch 1993), the Noah land surface model (Chen
and Dudhia 2001), the Rapid Radiative Transfer Model
for longwave radiation (RRTM;Mlawer et al. 1997), and
the Dudhia shortwave radiation scheme (Dudhia 1989).
The cumulus scheme is used only in the 27- and 9-km
domains.
a. Synoptic verification
1) THE FRONTAL SYSTEM
Sounding observations from the National Weather
Service (NWS) are compared with the simulated
temperature, geopotential height, and wind barbs on a
weather map at the 850-hPa pressure level at 0000 UTC
2 June 2008 (Fig. 2). The simulated geopotential heights
almost overlap with the observations (Fig. 2a), indicating
the front is well simulated. Over the frontal region, the
observed and simulated temperature fields are almost
identical. The larger temperature gradient over North
Dakota and the wind barbs representing a realistic
frontal systemwere reproduced by themodel simulation
(Fig. 2b).
2) THE LOW-LEVEL JET
To verify the forecast of the low-level jet, a comparison
is first made using NCEP Northern American Regional
FIG. 4. Time series of vertical profiles from 0000 UTC 1 Jun to 0000 UTC 3 Jun 2008 at Jayton (33.018N,
100.988W; elevation: 707 m): (a) wind speeds (contour interval is 5 m s21) and vectors obtained from NOAA
radar profiler and near-surface (10 m) winds from surface Mesonet observations, (b) wind speeds and vectors of
24–72-h model forecast from the 3-km domain, (c) temperatures obtained from NOAA radar profiler and near-
surface (2 m) temperature from surface Mesonet, and (d) temperatures of 24–72-h model forecast from the 3-km
domain. In (a) and (b), wind speeds greater than 10 m s21 are shaded. NOAA profiler data are not available
below 490 m AGL.
898 WEATHER AND FORECAST ING VOLUME 28
Reanalysis (NARR; Mesinger et al. 2006) data products.
Figure 3 compares wind speeds from the NARRdata and
the model simulations at 850 hPa at 0900 UTC 2 June
2008. The similarities between the wind fields in the
NARR and the simulation, in terms of jet coverage and
intensity, prove that the model has successfully simulated
the low-level jet.
Wind profile observations from the NOAA Profiler
Network in Jayton clearly reveal the nocturnal jet. A
time series of vertical profiles with wind speeds and
vectors from 0000 UTC 1 June to 0000 UTC 3 June 2008
are displayed in Fig. 4a. Compared with the observed
wind speeds and vectors, the simulated wind (Fig. 4b)
reasonably presents the structure of the low-level jets,
although the simulations underestimate the wind speed
and intensity of the low-level jet.
Meanwhile, surface winds (10-m wind from Mesonet
surface stations) are relatively calm during the night
because the turbulence ceases after sunset. However,
the simulated 10-m wind speeds during the night are
much higher than the observations, indicating the model
has not captured the decoupling between the surface
and higher-level air after sunset. Figure 4c shows that
the near-surface air cools quickly after sunset due to ra-
diative cooling. The air at 2 m is 38–58C colder than the
air above since radiative cooling begins at the surface.
The near-surface air temperature decreases immediately
after sunset and then decouples with the air in the re-
sidual layer above that stays relatively warmer during
the night. The simulated temperature (Fig. 4d) in the
boundary layer captures the observed temperature in-
versions on both nights in the higher-level air. However,
the simulated 2-m temperatures during both nights are
much warmer than the observations. As a result, the
model reproduces only a weak decoupling between the
surface and the air above.
b. Verification of near-surface atmosphericconditions
1) THE FRONTAL CASE OVER FLAT TERRAIN
Figure 5a shows the area and Mesonet observation
stations used for verifying the front. There are over 400
observations available hourly. Figure 5b compares the
realistic and model terrains in these stations. The model
terrain at all resolutions generally matches the actual
FIG. 5. The area and Mesonet observation stations used for verification: (a) for the front case and (c) for the low-
level jet case. (b),(d) Comparison of the actual andmodel terrain heights for stations in (a) and (c), respectively. D01,
D02, and D03 represent the domains at 27-, 9-, and 3-km horizontal resolutions, respectively. The straight dashed
lines in (b) and (d) denote Y 5 X.
JUNE 2013 ZHANG ET AL . 899
terrain, while the 3-km domain (D03) makes the best
match.
The simulated 2-m temperatures in the frontal area
generally agree well with observations in all domains
(Fig. 6a), showing no systematic bias. However, large
MAEs occur in 2-m temperature at the end of the
simulation when the stationary front changes to a cold
front (Fig. 6d). The MAEs of 2-m temperature are
smallest in the 3-km domain, specifically during the
second day.
The model captures the southwest-to-northeast wind
direction change accompanying the frontal passage (Fig.
6b). Particularly, wind directions in the simulation agree
well with the observations from 1200 UTC 1 June to
1200 UTC 2 June 2008. Relatively larger errors in wind
direction occur near the end of the simulation period
FIG. 6. Comparison of (a) 2-m temperature, (b) 10-m wind direction, and (c) 10-m wind speed between obser-
vations and simulations averaged over the area of the frontal system (e.g., the domain covered in Fig. 5a) and mean
absolute errors of simulated (d) 2-m temperature, (e) 10-m wind direction, and (f) 10-m wind speed. D01, D02, and
D03 represent results frommodel domains at horizontal resolutions of 27, 9, and 3 km, respectively. The shaded areas
indicate the nighttime, hereafter.
900 WEATHER AND FORECAST ING VOLUME 28
(Fig. 6e) and are caused mainly by the rapid transitions
from a cold front to a stationary front and then back
again to a cold front. These changes complicate the
frontal event and present additional difficulties for the
numerical simulations. Overall, the errors in 10-m
wind direction are similar in all three domains in the
first 36 h of forecasts. The 3-km domain performs
better during the last 12 h, when rapid transitions take
place. A diurnal feature of the errors in wind di-
rection, characterized by larger errors during night-
time and smaller errors during daytime, can also be
seen in Fig. 6e.
The observations clearly depict diurnal variations in
wind speeds, with higher speeds during the daytime and
lower speeds at night. The model well simulates the di-
urnal signals but generates larger errors at night (Figs. 6c
and 6f) during the nocturnal jet. In particular, the model
does well in simulating wind speeds between 0000 and
0200 UTC 1 June [corresponding to 1800–2000 central
standard time (CST) 31 May] and 1500 UTC 1 June and
0100 UTC 2 June (corresponding to 0900–1900 CST
1 June), both of these periods are during the daytime.
However, larger errors (Fig. 6f), characterized by
positive biases (Fig. 6c), are found for nocturnal wind
speeds. These positive bias errors in 10-m wind speed
can be attributed to the incomplete representation of
the decoupling between the higher-layer air and the
near-surface atmosphere in the simulation (similar to
that shown in Figs. 4b and 4d). In addition, the fore-
casts in the 3-km domain outperform the 9- and 27-km
domains during the daytime, but not at night. This is
mainly because the coarser-resolution domains do
not resolve the intensity of the low-level jets as well as
the higher-resolution domain does during the night-
time, and thus, the lower wind speeds produced by
the coarser-resolution domains have less impact on the
near-surface wind speeds. In other words, due to the
model’s inability to represent the decoupling between
the near-surface layer and the boundary layer above
(as mentioned in section 3a), the coarser-resolution
domains outperform the high-resolution domain dur-
ing nighttime.
2) THE LOW-LEVEL JET OVER FLAT TERRAIN
Figure 5c shows the Mesonet observation stations
used for verifying the low-level jet. Figure 5d compares
the actual and model terrain heights over these stations.
Themodel terrainmatches the actual terrain very well at
most stations. The terrain heights in the 3-km domain
(the innermost domain), again, best match the actual
terrain heights.
The 3-km domain results in the smallest errors in
2-m temperature during 1500 UTC 1 June–1200 UTC
2 June (Fig. 7a). However, it produces the largest errors
during the first night and errors that are comparable to
the other domains in the earlymorning for both days. The
FIG. 7. Mean absolute errors over the low-level jet area of sim-
ulated (a) 2-m temperature, (b) 10-m wind direction, and (c) 10-m
wind speed. D01, D02, and D03 represent results from model do-
mains at horizontal resolutions of 27, 9, and 3 km, respectively.
JUNE 2013 ZHANG ET AL . 901
errors peak at 1200 UTC 1 June and 1200 UTC 2 June,
when the boundary layers are most stable. Southerly flow
dominates during the simulation period. The errors in
10-m wind direction are relatively small due to the strong
southerly forcing (Fig. 7b). Owing to the influence of the
low-level jet and the inaccurate representation of the
nocturnal decoupling and radiative cooling as mentioned
above, there are relatively larger errors in 10-m wind
speed during the nighttime (Fig. 7c).
Accurate simulation of the transition boundary layer
and the typical stable boundary layer near the ground is
still one of the challenges in numerical simulation. The
method of parameterizing the stable boundary layer has
also been an active research area in recent studies (Brown
andWood 2003; Teixeira et al. 2008). A discussion of the
best way to overcome forecast errors in a stable boundary
layer is beyond the scope of this study.However, accurate
forecasts of near-surface conditions depend on the
model’s ability to simulate the stable boundary layer.
4. Simulation and verification: 1–3 December 2010
The simulation is initialized at 0000 UTC 31 November,
and the results from 1 to 3 December 2010 are used to
verify the persistent inversion. Three-level, one-way
nested domains (D01, D02, and D03) at 12-, 4-, and
1.33-km horizontal grid spacings (hereafter referred to
as the 12-, 4-, and 1.33-km domains, respectively) are
used. The innermost domain (i.e., the 1.33-km domain)
focuses on the Salt Lake Valley and its surrounding
mountains. The model also includes 37 vertical levels
from the surface up to 50 hPa.
Physical parameterization configurations for this case
(see Table 1) are the same as for the 1–3 June 2008 case,
except for the Mellor–Yamada–Janji�c (MYJ) turbulent
kinetic energy (TKE) PBL scheme (Mellor and Yamada
1982). The cumulus scheme is used only in the 12-km
domain.
a. Synoptic verification
Sounding observations, obtained every 12 h from the
station at the Salt Lake City International Airport
(KSLC) are available to examine the structure of the
atmospheric boundary layer. Model-simulated winds
and temperatures are interpolated to the sounding lo-
cations for comparison. Figure 8 shows the evaluation of
the temperature and wind fields for both soundings and
simulations throughout the 2-day period. Observations
show warmer air (relative to near-surface air) above
FIG. 8. Time series of vertical profiles of temperature
(contour interval is 18C) andwind at theKSLC (40.778N,
111.858W, elevation: 1289 m): (a) sounding observa-
tions, (b) ARW simulation from the 1.33-km domain
with 37 vertical levels, and (c)ARWsimulation from the
1.33-km domain with 70 vertical levels. Shaded contours
represent temperatures greater than 228C.
902 WEATHER AND FORECAST ING VOLUME 28
the surface; namely, an inversion layer was present
throughout the entire period, although the inversion was
more intense in the late stages. The top of the inversion
varies from 700 to 1200 m AGL during this period. The
simulation reproduces the persistent inversion for the
entire period. The wind shears, which were present as
southeasterly to southerly beneath the inversion layer
and southwesterly above it at 1200 UTC 2 December
and 0000 UTC 3 December 2010, are well captured in
the simulations, although the transition heights are
slightly different from the observations. The simulated
heights of the inversion layer are lower than those in the
sounding observations. The simulated temperature
gradient at the bottom of the boundary layer is not as
strong as the observed. These results are similar to those
of Hanna and Yang (2001). Apparently, discrepancies
between simulations and observations can be attributed
mainly to errors in the simulation of the near-surface
atmospheric conditions.
b. Verification of near-surface atmosphericconditions
Figure 9a shows the distribution of Mesonet obser-
vation stations used for verifying this case. In complex
terrain, observations are distributed unevenly in the Salt
Lake Valley and the surrounding mountains. Figure 9b
compares the model and actual terrain heights for all
three domains. The 1.33-km domain represents the
actual terrain substantially better than the 12- and
4-km domains. The 12-km domain misrepresents lower
(higher) terrain [less (greater) than 2000 m in the valley
(mountains)] with higher (lower) heights. Consequently,
the coarser-resolution domain does not resolve the deep
valley and sharp mountains accordingly. The 4-km do-
main has an intermediate ability to represent the terrain
compared with the 1.33- and 12-km domains.
Because of the large differences in terrain height be-
tween the stations in the valley and those in the moun-
tains, the stations are separated into two groups during
the verification: stations inside the Salt Lake Valley and
those in the surrounding mountains (valley stations and
mountain stations hereafter). A similar separation was
used in Hart et al. (2005).
1) SALT LAKE VALLEY: VALLEY STATIONS
Consistent with synoptic verification results, large
forecast errors are found for 2-m temperature at the
valley stations, especially during the night of 2 December
FIG. 9. As in Fig. 5, but for the persistent inversion case over the (a),(b) Salt Lake Valley and (c),(d) DPG. D01, D02,
and D03 represent the domains at 12-, 4-, and 1.33-km horizontal resolutions, respectively.
JUNE 2013 ZHANG ET AL . 903
2010 (Fig. 10a), when the valley was undergoing extreme
nocturnal cooling induced by the persistent inversion.
The 12-km domain produces the smallest MAEs during
that night, while the 4-km domain produces the largest
MAEs. This degradation accompanying the increase in
resolution from 12 to 4 km can be attributed to the
predictability of the intense inversion and the terrain
representation in the ARW model. Specifically, the
terrain heights in the 4-km domain are typically lower
than those in the 12-km domain at the valley stations.
During the inversions, the temperature usually increases
with height inside the valley. However, the model can-
not fully capture the intense cold pools, resulting in
a temperature profile that decreases with height in the
near-surface layer. Therefore, the 4-km domain pro-
duces warmer surface temperatures than the 12-km
FIG. 10. MAEs by station type for numerical simulations of (a),(b) 2-m temperature; (c),(d) 10-m wind direction;
and (e),(f) speed. Figures in the left column [(a), (c), and (e)] represent MAEs for valley stations, and these in the
right column [(b), (d), and (f)] represent mountain stations. D01, D02, and D03 represent the domains at 12-, 4-, and
1.33-km horizontal resolutions, respectively.
904 WEATHER AND FORECAST ING VOLUME 28
domain due to its deeper valley representation, causing
even larger warm biases. A similar argument was pre-
sented by Hart et al. (2005) using the MM5 model.
It is interesting that the MAEs of the 1.33-km domain
are intermediate between those of the 12- and 4-km do-
mains. As seen in Fig. 8, the 1.33-km domain captures
the inversion in the lower atmosphere and properly rep-
resents the temperature lapse rate (i.e., increase with
height). With a deeper topography and proper lapse rate
representation, the 1.33-km domain produces better 2-m
temperature forecasts than the 4-km domain. However,
MAEs of the 2-m temperature are still larger in the
1.33-km domain than in the 12-km domain because the
temperature lapse rate in the lower atmosphere (namely,
the inversion intensity) resolved by the 1.33-kmdomain is
much weaker than that of the sounding observations (as
shown in Fig. 8).
The TMAEs in wind direction over all the stations are
about 808 (Fig. 11b). Both the 4- and 1.33-km domains
produce degraded wind direction forecasts (Figs. 10c
and 11b) because of the contradiction between the
model’s failure to fully capture the strong inversion
and the better terrain representation in the higher-
resolution domains, as discussed above.
The MAEs of 10-m wind speed at the valley stations
are significantly reduced in the 4-km domain (Figs. 10e
and 11c). Much greater wind speeds are produced in
the 12-km domain because it has a shallower valley
floor (at a higher elevation than those of the 4- and
1.33-km domains). There are larger MAEs in wind
speed since the observed winds are relatively calm at
the surface.
2) SALT LAKE VALLEY: MOUNTAIN STATIONS
Meanwhile, increasing the model horizontal resolu-
tion from 12 to 4 and then 1.33 km substantially im-
proves the temperature forecasts over the mountain
stations (Figs. 10b and 11a). These forecasts benefit from
the sharper and better representation of mountain
terrain at the higher resolution while the mountain sta-
tions are above the cold pools.
The wind direction forecasts are improved in the 4-km
domain from the 12-km domain but are degraded in the
1.33-km domain from the 4-km domain (Fig. 10d). The
improvement in the 4-km domain is mainly because
the mountain stations are more connected to the free
atmosphere in the forecasts at higher resolution. The
reasons for the degradation in the 1.33-km domain are
unknown, reflecting the challenges in producing accu-
rate forecasts of near-surface atmospheric conditions.
In reality, the accurate forecasting of near-surface wind
direction is extremely difficult, as it depends not only on
the terrain representation in the model, but also on the
accurate prediction of wind speed, and the thermal and
dynamical forcings. In addition, low wind speeds at the
surface layer pose extra difficulties in forecasting wind
FIG. 11. Time-averaged MAEs of (a) 2-m temperature, (b) 10-m wind direction (degrees), and (c) 10-m wind speed by station type. D01,
D02, and D03 represent the simulations from the domains at 12-, 4-, and 1.33-km horizontal resolutions, respectively.
FIG. 12. MAEs of 2-m temperature at DPG. D01, D02, and D03
represent the simulations from the domains at 12-, 4-, and 1.33-km
horizontal resolutions, respectively.
JUNE 2013 ZHANG ET AL . 905
direction, although observations with speeds less than
1.5 m s21 are already excluded in the comparison.
The simulation in the 12-km domain produces the
smallest MAEs in wind speed. Slight degradations are
found in the forecasts of the 1.33-km domain and rela-
tively larger degradations are found in the 4-km domain
(Figs. 10f and 11c).
3) DUGWAY PROVING GROUND
The great complexity of the terrain in the Salt Lake
Valley presents a significant challenge to forecast veri-
fication. To further examine error characteristics in
near-surface variables in complex terrain, an additional
verification is conducted in theDugway ProvingGround
(DPG) area. DPG is located approximately 80 mi south-
west of Salt Lake City. It is characterized by complex
terrain and is surrounded on three sides by mountain
ranges. Considering the height of its mountains relative
to its flat regions, the DPG area is more representative
of common complex terrain features.
Figure 9c shows the DPG area map and stations used
for verification. Currently, there are a total of 31 auto-
matic surface stations in the DPG area. However, six
stations (denoted by circles) are not used for verification
in this case since they did not begin providing observa-
tions until 2011. Figure 9d compares the actual and
model terrain heights. Similar to the previous section,
the 1.33-km domain has the best terrain representation.
There is evidently a large error in both the 12- and 4-km
domains in representing the actual terrain at station
DPG16, located on a mountaintop at an elevation of
2149 m. For instance, the terrain height at DPG16 in the
12-km domain is only 1450 m.
The effect of nocturnal cooling in the DPG area is
shown in Fig. 12. The temperature error is as high as 98Cduring the cooling night. The 12- and 4-km domains
produce the smallest and the largest MAEs, respec-
tively, while the 1.33-km domain produces intermediate
MAEs. The 2-m temperature errors display similar
features to those of the valley stations, as seen in Figs.
10a and 11a, since they are under the same weather re-
gime and synoptic environment. The MAEs of wind
direction and speed also show patterns similar to those
of the valley stations (not shown). To eliminate the im-
pact of the mismatched terrain at DPG16, statistical
analyses are rerun but with this station excluded. The
results remain almost the same. This implies that the
representative error caused by terrain mismatch is not
the sole reason for the errors in simulated near-surface
variables in complex terrain.
5. Sensitivity to PBL schemes and verticalresolution
It has been recognized that PBL parameterization
schemes have a substantial influence on the simulation
of surface variables (Hu et al. 2010; Shin and Hong
2011). It is also commonly believed that simulations
could be improved with increased model vertical resolu-
tion, especially in the boundary layer, in which the model
can better resolve small-scale processes. To examine the
impact of PBL schemes and model vertical resolution on
the simulation of near-surface atmospheric conditions,
additional experiments are conducted and discussed in
this section.
a. Sensitivity to various PBL schemes
The ARW has multiple PBL scheme options, char-
acterized by different closure methods, prognostic
FIG. 13. TMAEs of (a) 2-m temperature, (b) 10-m wind direction,
and (c) 10-mwind speed for various cases and locations with various
PBL schemes. All results are from the innermost domain.
906 WEATHER AND FORECAST ING VOLUME 28
variables, cloud mixing, and other aspects. The first set
of sensitivity experiments uses various PBL schemes
while keeping other configurations the same, as specified
in the control simulations (Table 1). These experiments
are designed to evaluate the sensitivity of numerical
simulation of near-surface temperature and wind fields
to different PBL schemes.
Five PBL schemes in the ARW are tested and com-
pared—YSU, MYJ, quasi-normal scale elimination
(QNSE; Sukoriansky et al. 2005), Mellor–Yamada–
Nakanishi–Niino level 2.5 (MYNN2; Nakanishi and
Niino 2004), and the Asymmetric Convective Model
version 2 (ACM2; Pleim 2007). Among these schemes,
the YSU and ACM2 are first-order, nonlocal schemes.
They do not require additional prognostic equations to
describe the effects of turbulence on mean variables.
They are both based on the K profile in determining
the diffusivity in the boundary layer and consider
nonlocal mixing by convective large eddies. The YSU
scheme expresses nonlocal mixing by simply adding
a nonlocal gradient term in the eddy diffusion equa-
tion. The YSU scheme is modified in ARW version 3
(Hong and Kim 2008) to enhance mixing in the stable
boundary layer. The ACM2 scheme explicitly ex-
presses the nonlocal upward flux transport. The MYJ,
QNSE, and MYNN2 schemes are classified as 1.5-
order TKE closure schemes. They calculate eddy dif-
fusion coefficients by predicting TKE. They differ in
how they define the coefficients in the diffusion
equation. More details about these PBL schemes can
be found on the ARW Web site (http://www.mmm.
ucar.edu/wrf/users), as well as in Shin and Hong (2011)
and Hu et al. (2010).
Figure 13 depicts the sensitivity of near-surface vari-
able forecasts to model PBL schemes by comparing the
TMAEs of surface variables over the whole simulation
period in each case. For temperature (Fig. 13a), the
TMAEs of the cold front and the low-level jet in
summer 2008 are smaller than those of the inversion
case in winter 2010. For the frontal and low-level jet
cases, all schemes generate similar TMAEs. For the
inversion case, however, the TMAEs of temperature
are very sensitive to the choice of PBL scheme. The
MYJ and QNSE schemes produce smaller TMAEs
than the others at both valley and mountain stations.
All PBL schemes generate similar errors in wind di-
rection for all cases (Fig. 13b), except that the MYJ
and QNSE schemes produce relatively larger TMAEs
at the mountain stations. In terms of wind speed,
the ACM2 and MYNN2 schemes perform best in the
frontal and low-level jet events (Fig. 13c). For the
inversion, the YSU, MYNN2, and ACM2 schemes
perform equally well at both valley and mountain
stations. The MYJ and QNSE schemes, which pro-
duce the best temperature forecasts for the inversion
(as seen in Fig. 13a), are the least accurate in wind
speed forecasts.
Overall, all schemes perform similarly in the frontal
and low-level jet cases in terms of temperature and wind
direction forecasts. The MYJ and QNSE schemes pro-
duce relatively larger TMAEs in wind speed in the
frontal and low-level jet cases. They improve the tem-
perature forecasts at both valley and mountain stations
for the winter 2010 inversion but at the same time pro-
duce the least accurate wind speed and direction fore-
casts at the mountain stations.
Figure 14a further examines the sensitivity of 2-m
temperature simulation to model PBL schemes as
a function of forecast leading time in the Salt Lake
Valley for the December 2010 simulation. The five
schemes split into two groups. The group with the MYJ
and QNSE schemes performs much better than the
group with the three other schemes, especially during
the first day of the simulation. The MYJ and QNSE
schemes, however, have larger MAEs in the 10-m wind
speed forecast, as shown in Fig. 14b.
FIG. 14. MAEs of (a) 2-m temperature and (b) 10-m wind speed with various PBL schemes from the 1.33-km domain
for Salt Lake Valley stations.
JUNE 2013 ZHANG ET AL . 907
The above results indicate that no single PBL scheme
leads to overall improvement in the forecasts of near-
surface wind and temperature fields for all cases and
regions. The model is more sensitive to the choice of
PBL scheme in the inversion case in complex terrain
than in the frontal and low-level jet cases in flat terrain.
In addition, one PBL scheme that results in better wind
(temperature) forecasts does not reproduce improved
temperature (wind) forecasts. Because the MYJ and
YSU schemes have been frequently used in recent ap-
plications and also because of their respective perfor-
mance in accurately forecasting wind and temperature
fields, they were chosen for the control experiments for
the low-level jet in flat terrain and the inversions in
complex terrain, as listed in Table 1.
b. Sensitivity to model vertical resolution
The other set of experiments is used to examine the
sensitivity of model simulations to vertical resolution.
In these experiments, the vertical levels are increased to
70 (70L) instead of 37 (37L) as in the control simula-
tions. Figure 15 shows sketches of vertical model levels
in both configurations. The increase in the vertical levels
is more obvious below 6 km AGL, especially the lowest
2 km AGL. As seen in Fig. 8c, simulations with in-
creased vertical resolution can better resolve the struc-
ture of the persistent inversion in the boundary layer and
produce more reasonable inversion depth and intensity.
However, the higher vertical resolution does not im-
prove the forecasts of near-surface variables. Figure 16
shows that both experiments (with 37 and 70 vertical
levels) perform almost identically in terms of the TMAEs
for both 2-m temperature and 10-m wind speed and
direction for almost all cases and over all regions.
Overall, improvement in surface forecasts is not ensured
by increased vertical resolution, although it can help to
better resolve the structures of the mesoscale phenom-
ena in the boundary layer.
6. Characteristics of flow-dependent errors:Statistics over a 1-month period in the fall of 2011
The above results from the three typical weather
patterns indicate the case-by-case variability of the er-
rors in forecasts of near-surface variables. To further
understand the general characteristics of the errors in
near-surface forecasts in complex terrain, additional
verification is conducted for forecasts over a 1-month
period in the DPG area. We chose the DPG area as
the focus for multiple case statistics for two reasons: 1)
as mentioned above, the DPG area better represents
FIG. 15. Sketches of vertical models with (left) 37 and (right) 70
levels in AGL heights.
FIG. 16. As in Fig. 13, but for different model vertical resolutions.
908 WEATHER AND FORECAST ING VOLUME 28
common complex terrain features and 2) forecasts of
near-surface variables behave similarly over the DPG
area and the Salt Lake Valley (as seen in Fig. 12 and
described in section 4b), making it easier to examine the
forecast errors associated with synoptic forcings.
A near-real-time forecasting system was built us-
ing version 3.3 of the ARW. From 15 September to
14 October 2011, near-real-time forecasts were per-
formed 4 times daily (at 0000, 0600, 1200, and 1800UTC)
to produce a 48-h forecast each time. Four one-way
nested domains, with 30-, 10-, 3.33-, and 1.11-km hori-
zontal grid spacings, are used (Fig. 17). The innermost
domain (1.11 km) focuses on the DPG area. Initial
and boundary conditions are derived from the analy-
ses and forecasts produced by the NAM forecast sys-
tem at NCEP at 0000, 0600, 1200, and 1800 UTC. Over
a 1-month period, a total of 120 forecasts are gener-
ated. As shown in Table 1, in addition to using physical
schemes similar to those in the aforementioned case
study, the Purdue Lin microphysics scheme (Lin et al.
1983, Chen and Sun 2002) and the YSU PBL scheme
are used.
a. Overall evaluation
MAEs and BEs are employed to characterize the
forecast errors in near-surface variables. A statistical
calculation is done for each of the four initialization
times. For example, all forecasts initialized at 0000 UTC
during the month are averaged over all stations to
calculate the MAEs and BEs as a function of forecast
leading time.
Figures 18a–c show the MAEs calculated in the DPG
area in the 10-, 3.33-, and 1.11-km domains for the
forecasts initialized at 0000 UTC. Results confirm that
simulations at high resolution do not always outperform
those at coarser resolution. However, a clear diurnal
pattern is found in the errors of all variables produced by
all model domains. Specifically, the temperature error
peaks twice per day, around 0300 mountain standard
time (MST) and 1500 MST (corresponding to 1000
and 2200 UTC). There are also two error minima for
temperature at around 0700 and 1900 MST (corre-
sponding to 1400 and 0200 UTC). Wind speed and
direction follow the sameerror trends, with amaximum in
the early evening or before sunrise and a minimum in the
afternoon.
Using results from the 1.11-km domain, the depen-
dence of the surface forecasts on initialization time is
examined. Figures 18d–f show that the error trends are
independent of initialization time and forecast leading
time and follow the same diurnal variation. However,
compared with the forecasts initialized during the day-
time (0000 and 1800 UTC), relatively large errors occur
in the first 2–3 h in 2-m temperature for the forecasts
initialized at night (0600 and 1200 UTC). The large er-
rors in the nighttime-initiated forecasts could be caused
by the erroneous soil temperature initialization in the
NAM analysis. Apparently, the large errors associated
with initial conditions in the nighttime-initiated fore-
casts do not persist beyond a few hours. This may in-
dicate that a better local-scale initialization using data
assimilation can help reduce forecast errors within the
FIG. 17. Locations of model domains for near-real-time forecasting from 15 Sep to 14 Oct 2011.
JUNE 2013 ZHANG ET AL . 909
first few hours, but its impact may subsequently vanish.
Therefore, cycled data assimilation may be required to
mitigate the problem.
Overall, compared with the previous study by Liu
et al. (2008), the statistical errors are moderate for the
whole period, with maximum errors of 38C in 2-m
temperature and 2 m s21 in 10-m wind speed. Figure 19
further demonstrates the diurnal patterns of the bias errors
in 2-m temperature over the whole month. Positive
(warm) biases are found at night and negative (cold) biases
are present during the daytime. No systematic biases are
found in wind direction and speed (figures not shown).
FIG. 18. MAEs of simulated near-surface variables for (a)–(c) different model domains and (d)–(f) various
initialization times. Shown are the (a),(d) 2-m temperature; (b),(e)10-m wind speed; and (c),(f) 10-m wind di-
rection. Results in (a)–(c) are from 0000 UTC forecasts and D02, D03, and D04 represent results from model
domains at horizontal resolutions of 10-, 3.33-, and 1.11 km, respectively. Results in (d)–(f) are from the 1.11-km
domain and various curves represent forecasts initialized at different times. The forecasting period for all forecasts
is 48 h. The forecasts are output every 3 h in the 10-km domain (D02), and every hour in the 3.33- and 1.11-km
domains.
910 WEATHER AND FORECAST ING VOLUME 28
b. Strong versus weak synoptic forcing cases
In sections 3 and 4, it was concluded that the errors in
2-m temperature are significant under strong synoptic
forcing. With 1-month forecasts that include both qui-
escent periods and strong synoptic forcing cases, we
categorize the forecasts into strong and weak synoptic
forcing cases by checking the weather maps at the sur-
face and at the 700- and 850-hPa pressure levels. A
strong forcing case is identified when a cold front,
a closed low, a trough, a low pressure system, or large
wind speeds (greater than 5 m s21) is present at the
surface or on 700- or 850-hPa weather maps. In contrast,
a high pressure system, a ridge, low wind speeds (less
than 5 m s21) at 700 hPa or 850 hPa is identified as
a weak forcing case. Overall, three weak synoptic forc-
ing cases (i.e., 0000 UTC 21 September–1800 UTC 23
September, 0000 UTC 27 September–1800 UTC 29
September, and 0000 UTC 13 October–1800 UTC 15
October) and three strong forcing cases (i.e., 0000 UTC
16 September–1800 UTC 18 September, 0000 UTC
3October–1800UTC5October, and0000UTC5October–
1800UTC 7October) are identified. Figure 20 compares
the errors between the weak and strong forcing cases.
For each case, four forecasts with different initial times
are compared. It is apparent that diurnal patterns are
present in the forecast errors for the weak forcing ca-
ses. The errors are independent of initialization time
and forecast leading time (Figs. 20a,c,e). The strong
forcing cases, however, show flow-dependent features
(Figs. 20b,d,f). The forecast errors do not follow a diurnal
pattern, implying that the errors become more closely
related to the influence of the weather systems. In addi-
tion, the magnitude of the errors is generally greater in
the strong forcing than in the weak forcing cases.
7. Summary and concluding remarks
In this study, the performance of version 3.3 of the
ARW in predicting near-surface atmospheric temperature
and wind conditions under various terrain and weather
regimes is evaluated. Three individual events under
strong synoptic forcing, namely, a frontal system, a low-
level jet, and a persistent cold air pool, are first verified
against observations over both flat and complex terrain.
It is found that the ARW is able to produce reasonable
simulations of weather phenomena. Verification of
near-surface conditions (i.e., 2-m temperature and 10-m
wind) indicates the complexity in forecasting these sur-
face variables. For the frontal case and low-level jet case
over the central United States, the model terrain
matches the actual terrain and thus mitigates represen-
tative errors. The forecasts of surface variables generally
agree well with the observations. However, errors still
occur, depending on the model’s ability in forecasting the
structures in the lower-atmospheric boundary layer. For
the inversion case over the Salt Lake Valley, different
error characteristics are found over the mountain and
valley stations. Terrain mismatch and the ARW’s
limited ability to simulate near-surface atmospheric
conditions make the forecasting errors even more
complicated.
Overall, forecast errors in near-surface atmospheric
variables show flow-dependent features in all three
of the individual cases when strong synoptic forcings
are present. To better understand the characteristics
of flow-dependent errors in complex terrain as they
relate to near-surface forecasting, 1-month forecasts
(from 15 September to 14 October 2011) are con-
ducted over complex terrain in the western United
States. It is found that the forecast errors of surface
variables depend to a large degree on the diurnal cycle
of the surface variables themselves, especially when
the synoptic forcing is weak. The forecast errors for
2-m temperature reach two daily maxima at 0300 and
1500 local time, and two daily minima at 0700 and 1900
local time. Errors in wind speed and direction follow
the same trends, with a maximum at night and a mini-
mum in afternoon. Forecast errors follow the same
trends regardless of the initialization time, showing
that forecast errors are independent of the initializa-
tion time and forecast leading time. Further analyses
reveal positive (warm) temperature biases at night and
negative (cold) biases during the daytime. In contrast
to the 2-m temperature, wind direction and speed have
no systematic biases from a long-term perspective.
Under strong synoptic forcing, diurnal patterns in
forecast errors are broken, while flow-dependent er-
rors are clearly shown.
FIG. 19. Bias error of simulated 2-m temperature from the 1.11-km
domain with various initialization times. The forecasting period for
all forecasts is 48 h.
JUNE 2013 ZHANG ET AL . 911
Finally, it is apparent that simulations at finer reso-
lutions do not outperform those at coarser resolu-
tions in most cases. This was explained well in Hart
et al. (2005), who found that the inability of the
numerical model to depict near-surface structures
(such as strong temperature inversions) results in worse
forecasts, even with better terrain representation.
Meanwhile, increasing the model’s vertical resolution
FIG. 20. MAEs of simulated 2-m temperature from the 1.11-km domain for various cases: (a) 0000 UTC 21 Sep–
1800 UTC 23 Sep, (b) 0000 UTC 16 Sep–1800 UTC 18 Sep, (c) 0000 UTC 27 Sep–1800 UTC 29 Sep, (d) 0000 UTC
3 Oct–1800 UTC 5 Oct, (e) 0000 UTC 13 Oct–1800 UTC 15 Oct, and (f) 0000 UTC 5 Oct–1800 UTC 7 Oct. Four
forecasts with different initial times are available for each case. The forecasting period for all forecasts is 48 h.
Specifically, left panels [(a), (c), and (e)] represent weak synoptic forcing cases, and right panels [(b), (d), and (f)]
denote strong synoptic forcing cases.
912 WEATHER AND FORECAST ING VOLUME 28
does not help the predictability of near-surface variation,
although it improves the forecasts of mesoscale weather
phenomena. Numerical forecasts of near-surface atmo-
spheric conditions are also sensitive to the PBL scheme
in the ARW model, but there is no single PBL scheme
that performs better than the others. These factors illus-
trate the complexity and challenges involved in near-
surface simulation over complex terrain. Future work
should emphasize investigating the decoupling between
near-surface variables and the atmospheric boundary
layer and its impact on the predictability of near-surface
variables. The sensitivity of numerical predictions of
near-surface variables to terrain representation, land
surface parameters, and model errors will also need to be
examined.
Acknowledgments. The authors are grateful to the
NCAR WRF model development group. The Mesonet
observations used in this study were obtained online
(http://mesowest.utah.edu). The study was made pos-
sible in part due to the data made available by the
governmental agencies, commercial firms, and edu-
cational institutions participating in Mesowest. The
authors thank Mr. Christopher W. Pace, an under-
graduate student at the University of Utah, for his help
in preparing surface Mesonet data for this study.
Comments from Prof. Da-Lin Zhang and three anon-
ymous reviewers are highly appreciated, as they were
very helpful in improving an earlier version of the
manuscript.
This study is supported by Office of Naval Research
Award N00014-11-1-0709 and NSFGrants AGS 1243027
and 08339856. The computer time from the Center for
High Performance Computing (CHPC) at the University
of Utah is also gratefully acknowledged.
REFERENCES
Brown,A. R., andN.Wood, 2003: Properties and parameterization
of the stable boundary layer over moderate topography.
J. Atmos. Sci., 60, 2797–2808.
Chen, F., and J. Dudhia, 2001: Coupling an advanced land surface–
hydrology model with the Penn State–NCARMM5 modeling
system. Part I: Model implementation and sensitivity. Mon.
Wea. Rev., 129, 569–585.
Chen, S.-H., and W.-Y. Sun, 2002: A one-dimensional time
dependence cloud model. J. Meteor. Soc. Japan, 80, 99–
118.
Cheng, W. Y. Y., and W. J. Steenburgh, 2005: Evaluation of
surface sensible weather forecasts by the WRF and the Eta
Models over the western United States.Wea. Forecasting, 20,
812–821.
Dudhia, J., 1989: Numerical study of convection observed
during the Winter Monsoon Experiment using a meso-
scale two-dimensional model. J. Atmos. Sci., 46, 3077–
3107.
Hanna, S. R., and R. Yang, 2001: Evaluations of mesoscalemodels’
simulations of near-surface winds, temperature gradients, and
mixing depths. J. Appl. Meteor., 40, 1095–1104.
Hart, K. A., W. J. Steenburgh, and D. J. Onton, 2005: Model
forecast improvements with decreased horizontal grid
spacing over finescale intermountain orography during the
2002 Olympic Winter Games. Wea. Forecasting, 20, 558–
576.
Hong, S.-Y., and H.-L. Pan, 1996: Nonlocal boundary layer vertical
diffusion in a medium-range forecast model. Mon. Wea. Rev.,
124, 2322–2339.
——, and J.-O. J. Lim, 2006: The WRF single-moment 6-class
microphysics scheme (WSM6). J. Korean Meteor. Soc., 42,
129–151.
——, S.-W. Kim, J. Dudhia, M.-S. Koo, and K.-H. Seol, 2008:
Stable boundary layer mixing in a vertical diffusion packagee.
Ninth Annual WRF User’s Workshop, Boulder, CO, NCAR,
3.3. [Available online at http://www.mmm.ucar.edu/wrf/users/
workshops/WS2008/presentations/3-3.pdf.]
Horel, J., and Coauthors, 2002: Mesowest: Cooperative mesonets in
thewesternUnited States.Bull. Amer.Meteor. Soc., 83, 211–226.
Hu, X., J. Nielsen-Gammon, and F. Zhang, 2010: Evaluation of
three planetary boundary layer schemes in the WRF model.
J. Appl. Meteor. Climatol., 49, 1831–1844.
Kain, J. S., and J. M. Fritsch, 1993: The role of the convective
‘‘trigger function’’ in numerical prediction of mesoscale con-
vective systems. Meteor. Atmos. Phys., 49, 93–106.
Lee, T. J., R. A. Pielke, R. C. Kessler, and J. Weaver, 1989: In-
fluence of cold pools downstream of mountain barriers on
downslope winds and flushing. Mon. Wea. Rev., 117, 2041–
2058.
Leung, L. R., and Y. Qian, 2003: The sensitivity of precipitation
and snowpack simulations to model resolution via nesting in
regions of complex terrain. J. Hydrometeor., 4, 1025–1043.Lin, Y.-L., R. D. Farley, and H. D. Orville, 1983: Bulk parame-
terization of the snow field in a cloud model. J. Climate Appl.
Meteor., 22, 1065–1092.Liu, Y., and Coauthors, 2008: The operational mesogamma-scale
analysis and forecast system of the U.S. Army test and eval-
uation command. Part II: Interrange comparison of the ac-
curacy of model analyses and forecasts. J. Appl. Meteor.
Climatol., 47, 1093–1104.
Mass, C. F., D. Ovens, K. Westrick, and B. A. Colle, 2002: Does
increasing horizontal resolution produce more skillful fore-
casts? Bull. Amer. Meteor. Soc., 83, 407–430.Mellor, G. L., and T. Yamada, 1982: Development of a turbulence
closure model for geophysical fluid problems. Rev. Geophys.
Space Phys., 20, 851–875.
Mesinger, F., and Coauthors, 2006: North American Regional
Reanalysis. Bull. Amer. Meteor. Soc., 87, 343–360.
Mlawer, E. J., S. J. Taubman, P. D. Brown, M. J. Iacono, and S. A.
Clough, 1997: Radiative transfer for inhomogeneous atmo-
spheres: RRTM, a validated correlated-k model for the
longwave. J. Geophys. Res., 102 (D14), 16 663–16 682.
Nakanishi, M., and H. Niino, 2004: An improved Mellor–Yamada
level-3 model with condensation physics: Its design and veri-
fication. Bound.-Layer Meteor., 112, 1–31.
Pleim, J. E., 2007: A combined local and nonlocal closuremodel for
the atmospheric boundary layer. Part I: Model description and
testing. J. Appl. Meteor. Climatol., 46, 1383–1395.Rife, D. L., and C. A. Davis, 2005: Verification of temporal vari-
ations in mesoscale numerical wind forecasts.Mon.Wea. Rev.,
133, 3368–3381.
JUNE 2013 ZHANG ET AL . 913
Shafran, P. C., N. L. Seaman, andG.A. Gayno, 2000: Evaluation of
numerical predictions of boundary layer structure during the
Lake Michigan Ozone Study (LMOS). J. Appl. Meteor., 39,
412–426.
Shin, H. H., and S.-Y. Hong, 2011: Intercomparison of planetary
boundary-layer parameterizations in the WRF model for
a single day fromCASES-99.Bound.-LayerMeteor., 139, 261–
281.
Skamarock, W. C., and Coauthors, 2008: A description of the
Advanced Research WRF version 3. NCAR Tech. Note
NCAR/TN-4751STR, 113 pp.
Stull, R. B., 1988:An Introduction to Boundary LayerMeteorology.
Springer, 666 pp.
Sukoriansky, S., B. Galperin, and V. Perov, 2005: Application
of a new spectral theory of stably stratified turbulence to
atmospheric boundary layers over sea ice. Bound.-Layer
Meteor., 117, 231–257.
Teixeira, J., and Coauthors, 2008: Parameterization of the atmo-
spheric boundary layer: A view from just above the inversion.
Bull. Amer. Meteor. Soc., 89, 453–458.
Wolyn, P. G., and T. B. McKee, 1989: Deep stable layers in the
intermountain western United States. Mon. Wea. Rev., 117,
461–472.
Zhang, D.-L., andW. Zheng, 2004: Diurnal cycles of surface winds
and temperatures as simulated by five boundary-layer pa-
rameterizations. J. Appl. Meteor., 43, 157–169.
Zhong, S., and J. Fast, 2003: An evaluation of the MM5, RAMS,
and Meso-Eta models at subkilometer resolution using VTMX
field campaign data in the Salt Lake Valley. Mon. Wea. Rev.,
131, 1301–1322.
914 WEATHER AND FORECAST ING VOLUME 28